vulkan: temporary cary of vulkan fixes (#12971)

This should be reverted once we update ggml past b6897

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -120,81 +120,11 @@ shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
 
-#ifdef MUL_MAT_ID
-shared u16vec2 row_ids[BN];
-uint _ne1;
-
-#ifdef MUL_MAT_ID_USE_SUBGROUPS
-shared uvec4 ballots_sh[NUM_WARPS];
-
-void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
-    _ne1 = 0;
-    uint num_elements = p.nei1 * p.nei0;
-    uint nei0shift = findLSB(p.nei0);
-
-    uint ids[16];
-    uint iter = 0;
-
-    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
-        // prefetch up to 16 elements
-        if (iter == 0) {
-            [[unroll]] for (uint k = 0; k < 16; ++k) {
-                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
-                bool in_range = i < num_elements;
-                uint ii1;
-                if (nei0_is_pow2) {
-                    ii1 = i >> nei0shift;
-                } else {
-                    ii1 = i / p.nei0;
-                }
-                uint ii0 = i - ii1 * p.nei0;
-                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-            }
-        }
-        uint i = j + gl_LocalInvocationIndex;
-        bool in_range = i < num_elements;
-        uint ii1;
-        if (nei0_is_pow2) {
-            ii1 = i >> nei0shift;
-        } else {
-            ii1 = i / p.nei0;
-        }
-        uint ii0 = i - ii1 * p.nei0;
-        uint id = ids[iter++];
-        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-
-        ballots_sh[gl_SubgroupID] = ballot;
-        barrier();
-
-        uint subgroup_base = 0;
-        uint total = 0;
-        for (uint k = 0; k < gl_NumSubgroups; ++k) {
-            if (k == gl_SubgroupID) {
-                subgroup_base = total;
-            }
-            total += subgroupBallotBitCount(ballots_sh[k]);
-        }
-        barrier();
-
-        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
-        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
-            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
-        }
-        _ne1 += total;
-        iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
-            break;
-        }
-    }
-    barrier();
-}
-#endif // MUL_MAT_ID_USE_SUBGROUPS
-#endif // MUL_MAT_ID
-
 #ifdef COOPMAT
 shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
+#include "mul_mm_id_funcs.glsl"
 #include "mul_mm_funcs.glsl"
 
 void main() {

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -134,15 +134,15 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint ib = idx / 128;                         // 2 values per idx
             const uint iqs = idx % 128;                        // 0..127
 
-            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+            const uint qsi = (iqs / 64) * 16 + (iqs % 16);     // 0..15
             const uint scalesi = iqs / 8;                      // 0..15
             const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
 
-            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const uvec2 qs = uvec2(unpack8(data_a_packed16[ib].qs[qsi]));
             const uint scales = data_a[ib].scales[scalesi];
-            const vec2 d = vec2(data_a[ib].d);
+            const vec2 dm = vec2(data_a[ib].dm);
 
-            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+            const vec2 v = dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
 
             buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
 #elif defined(DATA_A_Q3_K)
@@ -179,7 +179,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint is = 2 * n + b;                 // 0..7
             const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
 
-            const vec2 loadd = vec2(data_a[ib].d);
+            const vec2 loadd = vec2(data_a[ib].dm);
 
             const uint scidx0 = (is < 4) ? is : (is + 4);
             const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -215,7 +215,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 
             const uint8_t hm = uint8_t(1 << (iqs / 16));
 
-            const vec2 loadd = vec2(data_a[ib].d);
+            const vec2 loadd = vec2(data_a[ib].dm);
 
             const uint scidx0 = (is < 4) ? is : (is + 4);
             const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -468,7 +468,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint ib = idx / 8;
             const uint iqs = (idx & 0x07) * 2;
 
-            const float d = e8m0_to_fp32(data_a[ib].e);
+            const float d = e8m0_to_fp32(data_a[ib].e) * 0.5;
             const uint vui = uint(data_a[ib].qs[iqs]);
             const uint vui2 = uint(data_a[ib].qs[iqs+1]);
 

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl

+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[BN];
+uint _ne1;
+
+#ifdef MUL_MAT_ID_USE_SUBGROUPS
+shared uvec4 ballots_sh[NUM_WARPS];
+
+void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
+            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
+        }
+        _ne1 += total;
+        iter &= 15;
+        if (_ne1 >= (ic + 1) * BN) {
+            break;
+        }
+    }
+    barrier();
+}
+#endif // MUL_MAT_ID_USE_SUBGROUPS
+#endif // MUL_MAT_ID

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl

+#if defined(DATA_A_Q4_0)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_Q4_1)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q5_0)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    uint32_t qh;
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_Q5_1)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[16/4];
+    uint32_t qh;
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q8_0)
+#define QUANT_R_MMQ 1
+// AMD likes 4, Intel likes 1 and Nvidia likes 2
+// #define BK_STEP 1
+struct block_a_cache {
+    int32_t qs[32/4];
+    FLOAT_TYPE dm;
+};
+#elif defined(DATA_A_MXFP4)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE d;
+};
+#elif defined(DATA_A_Q2_K)
+#define QUANT_R_MMQ 4
+struct block_a_cache {
+    uint32_t qs[2];
+    u8vec2 scales;
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q3_K)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[4];
+    FLOAT_TYPE_VEC2 d_scales;
+};
+#elif defined(DATA_A_Q4_K)
+#define QUANT_R_MMQ 2
+struct block_a_cache {
+    uint32_t qs[4];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q5_K)
+#define QUANT_R_MMQ 1
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 dm;
+};
+#elif defined(DATA_A_Q6_K)
+#define QUANT_R_MMQ 1
+struct block_a_cache {
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 d_scales;
+};
+#endif
+
+struct block_b_cache
+{
+    int32_t qs[8];
+    FLOAT_TYPE_VEC2 ds;
+};

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl

@@ -10,6 +10,7 @@ layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {int data_pos[];};
 layout (binding = 2) readonly buffer Z {float data_ff[];};
 layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+layout (binding = 4) readonly buffer I {uvec2 data_i[];}; // indices for set_rows
 
 layout (push_constant) uniform parameter {
     uint ncols;
@@ -27,6 +28,7 @@ layout (push_constant) uniform parameter {
     uint s2;
     int sections[4];
     uint is_back;
+    uint set_rows_stride;
 } p;
 
 float rope_yarn_ramp(const float low, const float high, const uint i0) {

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp

@@ -16,12 +16,19 @@ void main() {
     const uint row_x     = row_dst % ne1;
     const uint channel_x = row_dst / ne1;
 
-    const uint idst = row_dst*ne0 + i0/2;
+    uint idst = row_dst*ne0 + i0/2;
     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = row_x*ne0 + i0/2;
+        idst += data_i[channel_x].x * p.set_rows_stride;
+    }
+
     if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = data_a[ix + i0/2 + 0];
-        data_d[idst + i0/2 + 1] = data_a[ix + i0/2 + 1];
+        data_d[idst + i0/2 + 0] = D_TYPE(data_a[ix + i0/2 + 0]);
+        data_d[idst + i0/2 + 1] = D_TYPE(data_a[ix + i0/2 + 1]);
 
         return;
     }

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp

@@ -16,12 +16,19 @@ void main() {
     const uint row_x     = row_dst % ne1;
     const uint channel_x = row_dst / ne1;
 
-    const uint idst = row_dst*ne0 + i0;
+    uint idst = row_dst*ne0 + i0;
     const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
 
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = row_x*ne0 + i0;
+        idst += data_i[channel_x].x * p.set_rows_stride;
+    }
+
     if (i0 >= p.n_dims) {
-        data_d[idst + 0] = data_a[ix + 0];
-        data_d[idst + 1] = data_a[ix + 1];
+        data_d[idst + 0] = D_TYPE(data_a[ix + 0]);
+        data_d[idst + 1] = D_TYPE(data_a[ix + 1]);
 
         return;
     }

ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -566,7 +566,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && is_legacy_quant(tname)) {
+        // Integer dot mmq performs better with f32 accumulators
+        if (!f16acc && !coopmat && !coopmat2 && (is_legacy_quant(tname) || is_k_quant(tname) || tname == "mxfp4")) {
             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(merge_maps(base_dict, float_type_dict), {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
@@ -574,7 +575,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 }
 
 void process_shaders() {
-    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
 
     // matmul
     for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
@@ -841,10 +842,14 @@ void process_shaders() {
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
     string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});